9 – Training The Neural Network

The architecture of the net was very standard. It was taken from Google. It was a recurrent convolution. And our network is shown over here. Different layers. In fact, we experimented both with completely untrained networks and networks that had been pre-trained by googling other images. And then they would feed in the image that … Read more

8 – Solution Data Challenges

If you have the same intuition like me, you check the first three points, but not the fourth one. But it turns out, in our data, some dermatologists had taken specific instances of skin diseases and edit yellow markers into the images for sizing and the neural network could have perfectly classified that specific disease … Read more

7 – 07 Quiz Data Challenges V1

Collecting data from datasets, we have to clean up the data and integrate data, and if you become a serious machine learning engineer, you would spend a lot of time data cleanup. Here are four things in this quiz that one might find, and I want you to check mark the ones that you think … Read more

6 – 06 Image Challenge V3

Just to show how difficult it is to find skin cancer, I challenge you to look at this image, and if you have a doctor in your family or your friends, show them the image and ask them what they think is cancer. You can see different types of lesions, and you can see benign … Read more

5 – The Data

So let us talk about deep learning for skin cancer classification. My students, Brett and Andre, and our collaborators in the School of Medicine, collected roughly 130,000 images of skin conditions from various data sources, including the Web. Those include the images they are going to use for the competition. And these images came with … Read more

4 – Medical Classification

Here, you see the natural evolution of a melanoma over the years from 2003 to 2006. And there’s characteristics that doctors use for finding melanomas. They look at the fuzziness of the border, the asymmetry, the coloration, of course, the growth rate if it is accessible. But still, classifying melanomas, as I’ve shown you, is … Read more

3 – Survival Rate

As you have Googled, this is probably shocking to see how much difference is between stage zero and stage IV. I think the curves I’m showing you here at this moment show the difference in survival outlook for the different stages. What’s even more shocking is how similar stage zero and stage IV is. Look … Read more

27 – Mini Project Introduction

Back to skin images. We have, through an international collaboration of skin imaging, make available to you a dataset where you can now go and train your own skin classify using deep learning. The dataset is available next. And I hope you have fun building your own pocket dermatologist.

26 – Conclusion

Now we publish this result in the journal Nature. In fact, it became a cover of Nature. At the time, it was the fourth most discussed nature paper ever. And it really raised a lot of questions. What is the future of medicine look like if effectively, a camera-based system can outperform the best human … Read more

25 – Confusion Matrix

And it was also misclassify in an experiment where we looked at nine various classification for different types of disease, including cancers which is class seven and class eight. The bit of what’s called a confusion matrix. A confusion matrix is, suppose something is of class A, what is the probability of the network saying … Read more

24 – Confusion Matrix-Question 1

So after we develop a model, we want to find out how good it is. This is a difficult question. But in this section, we’ll learn a few different metrics that will tell us how good our model is. So we’re going to look at two main examples. The first example is a model that … Read more

23 – What Is The Neural Network Looking At

Finally, we ask the question, what is network fixating on? What is that you look at? And you can do a sensitivity analysis. You can wiggle your input image a little bit and see how much the output class of wiggles. And for some wiggles, if you wiggle them a little bit the output does … Read more

22 – Visualization

Machine Learning has beautiful techniques to visualize results. And one of the experiments we did is to take these incredibly high dimensional outputs and cluster them into two-dimensional space. These are technologies that you could learn now on machinery that we won’t go into detail but the image is fun to see. The East dot … Read more

21 – ROC Curve

Here’s our curve for Carcinoma. You find that in certain regimes, it’s 100 percent specific, 100 percent sensitive, and it’s only the little blue area on the top right that I are misclassifications. And there, we can now set thresholds to emphasize the detection of cancer or the reduce the cost of false diagnosis in … Read more

20 – Solution ROC Curve

And the answer is, if a network spits out a random number, then the output, the RC curve, is going to be a straight line. You can dial up your threshold. And the more dial up, the more cancers you’ll find, but there are also more non-cancerous you’ll find. And the ratio of cancerous-non-cancerous would … Read more

2 – 02 Skin Cancer V4

So, let’s dive in. To set the stage on, skin cancer is the most common cancer in the world. It’s not the most deadly but the most common. In the U.S., we find 5.4 million new cases of skin cancer every year. They come in different types. Some are called Carcinomas, some are called Melanomas. … Read more

19 – 17 Quiz ROC Curve 1 PT2 V1

So here is the ROC curve for our carcinoma classifier. In medicine, often we show the entire curve. You know them as ROC curve, the sensitivity specificity curve. And I’m going to have a quiz for you on which you should be able to answer. Suppose we made a neuron network that completely ignored the … Read more

17 – ROC Curve

Now we’ll learn another technique to evaluate a model called the receiver operator characteristic curve, or ROC curve for short. It works as follows; consider this data which is now one dimensional, so all the red and blue points lie in one line and we want to find the correct split. So, we can have … Read more

16 – 16 Solution Diagnosing Cancer V3

If you’ve chosen 0.5, then your chances of accidentally missing a cancer are exactly the same as your chances of calling something a cancer that isn’t. If you call it 0.8, you will be great in reducing the cost of diagnostics, but bad in helping people survive when they actually have skin cancer. 0.2 is … Read more

15 – 15 Quiz Diagnosing Cancer V3

Before we dive into the best strategy for diagnosing, let me just do a quiz. Have you ever know a network that outputs the probability of a melanoma? And we have now set a special value. When we decide to classify it as melanoma or as no melanoma. And their value is between zero and … Read more